Non-intrusive Load Monitoring Based On Switching Voltage

Transcript

Non-Intrusive Load Monitoring Based on Switching Voltage Transients and Wavelet Transforms Cesar Duarte, Paul Delmar, Keith W. Goossen, and Kenneth Barner Electrical and Computer Engineering Department University of Delaware Newark, DE, USA Eduardo Gomez-Luna Electrical Engineering Department Universidad del Valle, Grupo GRALTA Cali, Colombia IAC Student Meeting – Nov 1st and 2nd, 2012. Atlanta, GA Table of Contents • • • • • • • • • Introduction Applications (Motivation) Methods (Solution techniques) Non-Intrusive Load Monitoring (NILM) Systems Based on Switching Voltage Transients. Switching Voltage Transients Features Based on Transforms Feature Classification: Support Vector Machines (SVMs) Experimental Results Conclusion IAC Student Meeting – Nov 1st and 2nd, 2012. Atlanta, GA 1 INTRODUCTION Nonintrusive load monitoring (NILM) systems have been successfully proposed as a low cost method for monitoring of load profile, operations under faulted conditions and even human activity or behavior Fig. House Electrical diagram. Source: http://soleragroup.com/electrical-wiring-sunnyvale-what-is-a-circuit IAC Student Meeting – Nov 1st and 2nd, 2012. Atlanta, GA 2 APPLICATIONS (MOTIVATION) • Reduction of energy consumption (knowing detailed energy consumption) “…displaying only instantaneous power, to motivate savings of 5-15% … However, most solutions for obtaining appliance-specific feedback are expensive” M. E. Berges, H. S. Matthews, and L. Soibelman, “A System for Disaggregating Residential Electricity Consumption by Appliance” S. Darby, The effectiveness of feedback on energy consumption, Oxford, UK: Environmental Change Institute, University of Oxford, 2006. EPRI, Residential Electricity Use Feedback: A Research Synthesis and Economic Framework, Palo Alto, California: Electric Power Research Institue (EPRI), 2009. K. Ehrhardt-Martinez, k.A. Donnelly, and J.A. Laitner, “Advanced Metering Initiatives and Residential Feedback Programs: a Meta-Review for Household Electricity-Saving Opportunities,” Report E105, ACEEE 2010. 3 APPLICATIONS (MOTIVATION) • Collection of appliance end use data • Check the operation of load control systems. • Monitoring human activity. • Schedule and health of loads on shipborads. IAC Student Meeting – Nov 1st and 2nd, 2012. Atlanta, GA 4 APPLICATIONS (COMPANIES) • Intel: Prototype appliance signature detection products. 2010. IAC Student Meeting – Nov 1st and 2nd, 2012. Atlanta, GA 5 APPLICATIONS (COMPANIES) • Belkin (Bought Zensi in 2010) IAC Student Meeting – Nov 1st and 2nd, 2012. Atlanta, GA 6 APPLICATIONS (COMPANIES) • GE (Cognitive Electric Power Meter) IAC Student Meeting – Nov 1st and 2nd, 2012. Atlanta, GA 7 APPLICATIONS (COMPANIES) • IBM (Watzzup System) IAC Student Meeting – Nov 1st and 2nd, 2012. Atlanta, GA 8 APPLICATIONS (COMPANIES) • Navetas (UK) IAC Student Meeting – Nov 1st and 2nd, 2012. Atlanta, GA 9 APPLICATIONS (COMPANIES) • Enetics (Associated to G. Hart): SPEED : Single Point End-Use Energy Disaggregation). IAC Student Meeting – Nov 1st and 2nd, 2012. Atlanta, GA 10 APPLICATIONS (COMPANIES) • 4home (Acquired by Motorola) IAC Student Meeting – Nov 1st and 2nd, 2012. Atlanta, GA 11 APPLICATIONS (COMPANIES) • Verlitics (Formerly Emme) IAC Student Meeting – Nov 1st and 2nd, 2012. Atlanta, GA 12 METHODS • Active and Reactive Power IAC Student Meeting – Nov 1st and 2nd, 2012. Atlanta, GA 13 METHODS • Power Transients • Current Harmonics IAC Student Meeting – Nov 1st and 2nd, 2012. Atlanta, GA 14 METHODS • V-I curves IAC Student Meeting – Nov 1st and 2nd, 2012. Atlanta, GA 15 METHODS • Switching Voltage Transients Off-On. Continuous Voltage Noise Plot of fan transition On-Off. 2 1 0 -1 6 2.8 67s] on Off-On. 3 4 4.2 4.4 4.6 4.8 Time [1 = 0.016667s] Plot of blender transition On-Off. 5 1 0 -1 4.5 67s] ansition Off-On. 67s] • 2.2 2.4 2.6 2.8 3 Time [1 = 0.016667s] Plot of incandescent lamp transition On-Off. 3.2 1 0 -1 -2 3.8 4 3 3.2 3.4 3.6 Time [1 = 0.016667s] 3.8 4 IAC Student Meeting – Nov 1st and 2nd, 2012. Atlanta, GA 16 METHODS Steady - State Low Frequency P, Q, Power Factor Admittance V-I curves High Frequency Current Harmonics V-I curves Continuous Noise (SMPS) Transient - State Power Transients Switching Voltage Transients Hybrid Methods Fig. House Electrical diagram. Source: http://soleragroup.com/electrical-wiring-sunnyvale-what-is-a-circuit IAC Student Meeting – Nov 1st and 2nd, 2012. Atlanta, GA 17 PCC Load Non-Intrusive Load Monitoring (NILM) Systems Based on Switching Voltage Transients Zs Vs H1(s) VM H2(s) (a) PCC Load Zs Vs H1(s) H2(s) (b) PCC VM Load Load Z2 H1(s) Zs Vs Vs H2(s) VM Z1 VM (c) IAC Student Meeting – Nov 1st and 2nd, 2012. Atlanta, GA 18 Switching Voltage Transients • Non-ideal movement of the contacts (e.g. bouncing) • Non-ideal conduction of the air gap (i.e. arcing) R2 500 Vr L2 0 -500 R1 VM V1 C1 L1 Voltage (V) C2 Vs Simulated Transient Recovery Voltage: Vr. -1000 -1500 -2000 -2500 -3000 0 0.1 0.2 0.3 Time (ms) 0.4 IAC Student Meeting – Nov 1st and 2nd, 2012. Atlanta, GA 0.5 19 Switching Voltage Transients Plot of fan transition Off-On. Plot of fan transition On-Off. 2 2 1 1 0 0 -1 -1 2 2.2 2.4 2.6 2.8 Time [1 = 0.016667s] Plot of blender transition Off-On. 3 4 1 1 0 0 -1 -1 3.5 4 4.5 Time [1 = 0.016667s] Plot of incandescent lamp transition Off-On. 2.2 4.2 4.4 4.6 4.8 Time [1 = 0.016667s] Plot of blender transition On-Off. 5 2.4 2.6 2.8 3 Time [1 = 0.016667s] Plot of incandescent lamp transition On-Off. 3.2 1 1 0 0 -1 -1 -2 3 3.2 3.4 3.6 Time [1 = 0.016667s] 3.8 4 3 3.2 3.4 3.6 Time [1 = 0.016667s] 3.8 IAC Student Meeting – Nov 1st and 2nd, 2012. Atlanta, GA 4 20 Switching Voltage Transients Plot of fan trans ition Off-On. Plot of fan trans ition On-Off. 2 2 1.8 1.5 1.6 1 1.4 1.2 0.5 1 0 0.8 0.6 -0.5 0.4 0 20 40 60 80 Time [1 = 1e-006s ] 100 120 0 50 Plot of blender transition Off-On. 100 150 200 250 Time [1 = 1e-006s ] 300 350 400 Plot of blender transition On-Off. 0.5 1.6 1.4 0 1.2 1 -0.5 0.8 0.6 -1 0.4 0.2 -1.5 0 -2 0 2 Plot 4 6 8 Time [1 = 1e-006s] of inc andes c ent 10 12 -0.2 0 lamp trans ition Off-On. 2 Plot -0.5 4 6 8 Time [1 = 1e-006s] of inc andes c ent 10 12 14 lamp trans ition On-Off. -1.65 -0.55 -1.7 -0.6 -1.75 -0.65 -1.8 -0.7 -1.85 -0.75 -1.9 -0.8 -1.95 -0.85 -2 -0.9 -2.05 -0.95 0 50 100 Time [1 = 1e-006 s ] 150 200 0 5 10 15 20 25 Time [1 = 1e-006 s ] 30 35 40 21 Switching Voltage Transients Voltage VM for a blender connection. Voltage (V) 200 0 -100 0 b) Initial conditions estimated to reduce natural response Voltage (V) Voltage (V) 100 100 a) Initial conditions equal to zero filter natural response Switching transient -100 Switching transient -200 0 5 Continuous noise 10 15 Time (ms) 20 25 Switching transient 100 0 -100 0 10 20 30 Time (ms) 40 50 A series of rapid closures and openings IAC Student Meeting – Nov 1st and 2nd, 2012. Atlanta, GA 22 Features Based on Transforms Set of signals TRANSFORMS (STFT, Wavelet, …) FEATURE EXTRACTION: Energy, SD Feature Vectors (M) SVM C.V. ACCURACY IAC Student Meeting – Nov 1st and 2nd, 2012. Atlanta, GA 23 Features Based on Transforms • Short Time Fourier Transform – STFT 𝒗= -50 -100 FFT 2048 Voltage (V) 0 1 𝑁 𝑁 𝐹𝐹𝑇𝑖2048 ,𝑁 𝑖=1 N 1µs windows -150 495 500 Time (s) 505 IAC Student Meeting – Nov 1st and 2nd, 2012. Atlanta, GA 24 Features Based on Transforms • Continuous Wavelet Transform 1 0.8 0.6 0.4 0.2 0 2 2.2 Wx = sqrt(Ts) * cwt( x, a/Ts, 'cmor14.0461-2.605' ) 2.4 2.6 Frequency 2.8 3 25 Features Based on Transforms Volts (V) Continuous Wavelet Transform Analysed Signal 200 0 -200 Continuous Wavelet Transform. Wavelet: cmor14.0461-2.605 35 30 Scales (a) • 25 20 15 10 5 0 20 40 60 time (s) IAC Student Meeting – Nov 1st and 2nd, 2012. Atlanta, GA 80 26 Features Based on Transforms • Wavelet Mother: Morlet 1 0.8 … 0.6 0.4 0.2 0 2 2.2 2.4 2.6 Frequency 2.8 3 27 Features Based on Transforms • Feature Vector Analysed Signal Volts (V) 100 0 -100 Continuous Wavelet Transform. Wavelet: cmor14.0461-2.605 35 Scales (a) 30 25 20 15 10 5 0 2 4 6 8 10 time (s) 12 14 16 Wx = sqrt(Ts) * cwt ( x, a/Ts, 'cmor14.0461-2.605' ) 28 Feature Classification: Support Vector Machines (SVMs) 1 𝑇 min 𝑤 𝑤 + 𝐶 2 𝐿 𝜉𝑖 𝑖=1 gn gn-1 Support Vectors ….. g3 g2 g1 C1 C2 C3 𝐾(𝒗𝑖 , 𝒗𝑗 ) = 𝑒 −𝛾 𝒗𝑖 −𝒗𝑗 IAC Student Meeting – Nov 1st and 2nd, 2012. Atlanta, GA ……… Cn-1Cn 2 29 Feature Classification: Support Vector Machines (SVMs) • K-fold Cross Validation 1 Fold K - 1 Folds 1 Testing Training 2 Training Testing ... K ... Training ... Training Training ... Training ... Training Testing LIBSVM libraries available at http://www.csie.ntu.edu.tw/~cjlin/libsvm. Chih-Chung Chang and Chih-Jen Lin, “LIBSVM: a library for support vector machines,” ACM Transactions on Intelligent Systems and Technology, 30 vol. 2, pp. 2:27:1–27:27, 2011 Experimental Results Appliance To wall socket Digital Oscilloscope Appliance Description VM Power Rating Desk Fan 48 VA Single Serve Blender 200 W Incandescent Desk Lamp 60 W Type of Switch Rotary sliding contacts. Both rotation ways. States: Off, speed II and speed I Normally open push to make. States: On and Off Rotary sliding contacts. Only clockwise rotation. States: On and Off. IAC Student Meeting – Nov 1st and 2nd, 2012. Atlanta, GA 31 Experimental Results Appliance Description Desk Fan Single Serve Blender Incandescent Lamp TOTAL Connection 16 8 13 37 Disconnection 15 8 10 33 IAC Student Meeting – Nov 1st and 2nd, 2012. Atlanta, GA TOTAL 31 16 23 70 32 Experimental Results 200V c) 0V -100V 0s 50V e) 0V -100V 50s 100s 150s 0s 200V 0V -100V 50s 100s 150s 0s 20V 0V 0V 0s -20V 50s 100s 150s 0s b) Fan 200V 300s 600s 900s d) Blender 0V -100V 0s Inc. Lamp Blender Fan 200V a) Disconnection 50s 100s 150s f) Inc. Lamp Connection 50s 100s 150s IAC Student Meeting – Nov 1st and 2nd, 2012. Atlanta, GA 33 Classification Accuracy Approach STFT CWT. Complex Morlet Wavelet Accuracy (10-fold) Best C Best g 71.43 % 831.7465 0.125 Feature vector size 2049 80 % 7.054106 0.3923 98 IAC Student Meeting – Nov 1st and 2nd, 2012. Atlanta, GA 34 Conclusion We have shown here that the use of a wavelet transform to classify results in more accurate results, compared to previous Fourier transform techniques. More importantly, it reduces the required vector size by over an order of magnitude, thus substantially lowering the computational requirements of the system. It can be expected that NILM systems will employ the classification methods outlined in this paper IAC Student Meeting – Nov 1st and 2nd, 2012. Atlanta, GA 35 THANKS! Questions/Comments IAC Student Meeting – Nov 1st and 2nd, 2012. Atlanta, GA 36